Skip to content
The Injection_
LIVE
LIVElast sweep next in--:--:--

[ RSS ][ JSON ][ llms.txt ]

INFOTool

lend: a broker that keeps agent credentials on the laptop

lend lets a remote coding agent borrow the laptop's commands and SSH keys one approved request at a time, rather than holding copies.

· Antonio Mika · Research by Antonio Mika

AI coding agents
INFOTool

lend is an open-source Go tool that lets a coding agent on a remote host ask the user's laptop to run allowlisted commands and sign with named SSH keys, each request approved locally instead of copied onto the host.

01What happened

Antonio Mika released lend on November 13, 2026, alongside a write-up on why coding agents running on remote VMs should not hold the user's credentials. The post sets out the two usual options and their costs: copying credentials onto the VM, where any process can read them, or forwarding the SSH agent, which lets the VM use every key with no approval and no record. lend reached the Hacker News front page on November 17.

02How it works

A policy-checked proxy over the forwarded SSH agent socket is the mechanism. The laptop runs lend agent with a JSON policy, the remote host reaches it through ssh -A, and the agent on the host calls lend exec or an MCP tool. The policy denies unlisted commands, SSH keys and agent management operations by default, and each admitted request still needs local approval, which can be device authentication, a prompt or a terminal confirmation. Only keys named by exact fingerprint are visible or signable.

03Who is affected

Developers who run coding agents on remote VMs or shared hosts are the audience, on macOS and Linux. The README describes lend as an initial implementation and is explicit about its limits: approval is not an execution sandbox, so an approved program can still read local files, credentials and environment variables. Caller details that a remote process reports are informational, because the README says a remote process can spoof them.

04What to do

Install lend through Homebrew or its checksum-verified script, then write a policy listing only the commands and key fingerprints an agent needs. Avoid the none approval mode, which the README warns lets anyone with socket access act without a prompt. Treat approving a shell or interpreter as authorizing arbitrary execution, because the approval prompt does not review stdin.

Sources 3 links, all fetched

  1. 1blog.antonio.cxLaunch Post
  2. 2github.comlend Repository
  3. 3lend.antonio.cxlend Documentation